Armando, I'm afraid the information you have provided is not sufficient to give meaningful answers/suggestions. What exactly do you mean when you say you don't know what products your customers bought? if you mean you don't have meta-data about each product your customers bought, that's not a problem as long as when two different customers purchase the same product (no matter what it is), your transactions log can ascertain this fact, i.e. you have a database of transactions of the form . (I'm sure you know that most CF algorithms, or SVD-based etc. work on (userid,itemid) pairs, and provide quite good results). 

If this fact cannot be established, and you don't have any meta-data for each product the users purchase, then I don't see how you can come up with any approach at all that can be expected to work. Neither collaborative-filtering techniques, nor clustering-based, nor algebra-based, nor content-based will help if neither of the two following conditions hold:

1. some kind of external information (meta-data) can be established about the items.

2. when two different transactions contain the same item purchased, this fact can be inferred from the recorded data.

As far as the "broad range" of products is concerned, this should not be a problem with a big-enough data-set. If your data-set comes from a credit-card company, I would expect your data-set can be as large as you like, and thus the "product variety" problem will not really matter, even though, knowing product hierarchies would be beneficial, and could further boost performance.

You might have to use a combination of techniques here:

Purchase propensity can in principle be predicted by credit card usage frequency. You can use customer binning to classify his propensity. The next step would be to use O-Joun Lee suggestion on clustering products.This suggestion might be your best hope since generalization of the categories could reduce your error rate. From there you could use a labeled data set to try to match both profiles.

just a thought

Armando, I'm afraid the information you have provided is not sufficient to give meaningful answers/suggestions. What exactly do you mean when you say you don't know what products your customers bought? if you mean you don't have meta-data about each product your customers bought, that's not a problem as long as when two different customers purchase the same product (no matter what it is), your transactions log can ascertain this fact, i.e. you have a database of transactions of the form . (I'm sure you know that most CF algorithms, or SVD-based etc. work on (userid,itemid) pairs, and provide quite good results). 

If this fact cannot be established, and you don't have any meta-data for each product the users purchase, then I don't see how you can come up with any approach at all that can be expected to work. Neither collaborative-filtering techniques, nor clustering-based, nor algebra-based, nor content-based will help if neither of the two following conditions hold:

1. some kind of external information (meta-data) can be established about the items.

2. when two different transactions contain the same item purchased, this fact can be inferred from the recorded data.

As far as the "broad range" of products is concerned, this should not be a problem with a big-enough data-set. If your data-set comes from a credit-card company, I would expect your data-set can be as large as you like, and thus the "product variety" problem will not really matter, even though, knowing product hierarchies would be beneficial, and could further boost performance.

Armando, please specify, what exactly do you know about the products from your data.

Without knowing something about products or at least having a possibility to establish a link between a transaction and a product  you are unable to provide recommendations of other products (and you can only recommend transactions). 

Thanks for all the pertinent feedback. 

Concerning the products, at this point I only have the names of the merchandiser, value and date of transactions.

We create some product categories using heuristic rules based on the name of the merchandiser and the activity code.

My idea is to use some content based RS using enriched data obtained from merchandiser website.

Dear Armando, given your latest descriptions, I am afraid it will be a very difficult task to provide meaningful product recommendations that would yield recall-at-10 rates of even 0.1. At the level of aggregation your information is, the best you can do is probably recommend individual merchandisers to the credit card customers, which may not be that bad after all. It is quite likely you can use the money your customers spend as "ratings" for your merchandisers awarded by your customers.

When you have some technical report ready, and you make it public, please let me know so I can read it as well. I'd be very interested in the solution you come up with...

Yes, that is what we are planning to do. A platform to recommend customers to merchandisers based on spending behaviour + location + socio-demographics. This seems to be a very classic problem but did not found any adequate approach. Right now, want I'm using is a classifier, for each category, to predict the probability of buying a product. This seems a little silly, I know. But it was the only approach I envisaged. It works well but it doesnt scale.